open-problem

Status: open (2026-10-24). Area: variational inference, optimisation · Prerequisites: statistical games, the envelope theorem, Fisher’s identity · Related: Factors are Parameterized Statistical Games, L-SPEC-002 Stop Gradient, the CT-ML wiki’s Statistical Game and Automatic Differentiation of Expected Values.

Sources: St Clere Smithe & Perin, AutoBayes, arXiv:2503.18608, 2025 (Definitions 27–29, Theorem 23, Remarks 24, 30); the question is original to this vault.

Bibliography: St Clere Smithe & Perin 2025

Question

When parameterized statistical games are composed, AutoBayes composes their gradients locally (Definition 29) and drops two terms: where an upstream parameter moves the pushforward prior seen downstream, and where a downstream parameter moves the sampling distribution of an upstream inversion (Remark 30). When are the dropped terms zero, and when they are not, how far is the fixed point of local-gradient descent from a stationary point of the true composite free energy?

Setting

Two games with parameter and with parameter , a prior on (as a prior game, Remark 24), an observation , and the composite free energy of Theorem 23. The local gradient is with the inversions held fixed. In Lenticulum this is the intended meaning of DiagonalCoupling(), the default annotation on every edge. The annotation does not implement AD stops or a graph backward pass; non-default annotations are refused at construction (LP-011 option 2).

What is known

  • Expected, to be proved: if every inversion is exact (the Bayesian posterior), both dropped terms vanish. The term through the inversion vanishes by the envelope theorem, because the free energy is stationary in the belief at the exact posterior. The term through the pushforward prior should vanish by Fisher’s identity . Then local gradients are exact gradients of , as in the EM algorithm.
  • With amortised or otherwise approximate inversions the terms are nonzero in general; this is what pathwise (reparametrisation) and score-function estimators restore, at a cost in variance.

What would count as an answer

A proof of the exact-inversion case; for approximate inversions, a bound on the distance between the fixed points of local and true gradient descent in terms of the inversion error (for example the KL divergence from the exact posterior); and a linear-Gaussian chain where both can be computed in closed form, as a test.

Why it matters here

It decides when DiagonalCoupling can stay the default and when a graph must ask for PathwiseCoupling or ScoreFunctionCoupling, which the stop-gradient specification (L-SPEC-002 Stop Gradient) has to settle.

Leads

The envelope theorem for the exact case; the variational EM literature for the approximate one (the gap between the ELBO gradient and the evidence gradient is a KL derivative).